如何使用 pyspark 在 Spark 2.0 中构建 sparkSession？

Question

I just got access to spark 2.0; I have been using spark 1.6.1 up until this point. Can someone please help me set up a sparkSession using pyspark (python)? I know that the scala examples available online are similar ( here ), but I was hoping for a direct walkthrough in python language.

My specific case: I am loading in avro files from S3 in a zeppelin spark notebook. Then building df's and running various pyspark & sql queries off of them. All of my old queries use sqlContext. I know this is poor practice, but I started my notebook with

sqlContext = SparkSession.builder.enableHiveSupport().getOrCreate() .

I can read in the avros with

mydata = sqlContext.read.format("com.databricks.spark.avro").load("s3:...

and build dataframes with no issues. But once I start querying the dataframes/temp tables, I keep getting the "java.lang.NullPointerException" error. I think that is indicative of a translational error (eg old queries worked in 1.6.1 but need to be tweaked for 2.0). The error occurs regardless of query type. So I am assuming

1.) the sqlContext alias is a bad idea

and

2.) I need to properly set up a sparkSession.

So if someone could show me how this is done, or perhaps explain the discrepancies they know of between the different versions of spark, I would greatly appreciate it. Please let me know if I need to elaborate on this question. I apologize if it is convoluted.

Answer 1

from pyspark.sql import SparkSession
spark = SparkSession.builder.appName('abc').getOrCreate()

now to import some .csv file you can use

df=spark.read.csv('filename.csv',header=True)

Answer 2

As you can see in the scala example, Spark Session is part of sql module. Similar in python. hence, seepyspark sql module documentation

class pyspark.sql.SparkSession(sparkContext, jsparkSession=None) The entry point to programming Spark with the Dataset and DataFrame API. A SparkSession can be used create DataFrame, register DataFrame as tables, execute SQL over tables, cache tables, and read parquet files. To create a SparkSession, use the following builder pattern:

>>> spark = SparkSession.builder \
...     .master("local") \
...     .appName("Word Count") \
...     .config("spark.some.config.option", "some-value") \
...     .getOrCreate()

Answer 3

From here http://spark.apache.org/docs/2.0.0/api/python/pyspark.sql.html
You can create a spark session using this:

>>> from pyspark.sql import SparkSession
>>> from pyspark.conf import SparkConf
>>> c = SparkConf()
>>> SparkSession.builder.config(conf=c)

Answer 4

spark  = SparkSession.builder\
                  .master("local")\
                  .enableHiveSupport()\
                  .getOrCreate()

spark.conf.set("spark.executor.memory", '8g')
spark.conf.set('spark.executor.cores', '3')
spark.conf.set('spark.cores.max', '3')
spark.conf.set("spark.driver.memory",'8g')
sc = spark.sparkContext

Answer 5

Here's a useful Python SparkSession class I developed:

#!/bin/python
# -*- coding: utf-8 -*-

######################
# SparkSession class #
######################
class SparkSession:

    # - Notes:
    # The main object if Spark Context ('sc' object).
    # All new Spark sessions ('spark' objects) are sharing the same underlying Spark context ('sc' object) into the same JVM,
    # but for each Spark context the temporary tables and registered functions are isolated.
    # You can't create a new Spark Context into another JVM by using 'sc = SparkContext(conf)',
    # but it's possible to create several Spark Contexts into the same JVM by specifying 'spark.driver.allowMultipleContexts' to true (not recommended).
    # - See:
    # https://medium.com/@achilleus/spark-session-10d0d66d1d24
    # https://stackoverflow.com/questions/47723761/how-many-sparksessions-can-a-single-application-have
    # https://stackoverflow.com/questions/34879414/multiple-sparkcontext-detected-in-the-same-jvm
    # https://stackoverflow.com/questions/39780792/how-to-build-a-sparksession-in-spark-2-0-using-pyspark
    # https://stackoverflow.com/questions/47813646/sparkcontext-getorcreate-purpose?noredirect=1&lq=1

    from pyspark.sql import SparkSession

    spark = None   # The Spark Session
    sc = None      # The Spark Context
    scConf = None  # The Spark Context conf

    def _init(self):
        self.sc = self.spark.sparkContext
        self.scConf = self.sc.getConf() # or self.scConf = self.spark.sparkContext._conf

    # Return the current Spark Session (singleton), otherwise create a new oneÒ
    def getOrCreateSparkSession(self, master=None, appName=None, config=None, enableHiveSupport=False):
        cmd = "self.SparkSession.builder"
        if (master != None): cmd += ".master(" + master + ")"
        if (appName != None): cmd += ".appName(" + appName + ")"
        if (config != None): cmd += ".config(" + config + ")"
        if (enableHiveSupport == True): cmd += ".enableHiveSupport()"
        cmd += ".getOrCreate()"
        self.spark = eval(cmd)
        self._init()
        return self.spark

    # Return the current Spark Context (singleton), otherwise create a new one via getOrCreateSparkSession()
    def getOrCreateSparkContext(self, master=None, appName=None, config=None, enableHiveSupport=False):
        self.getOrCreateSparkSession(master, appName, config, enableHiveSupport)
        return self.sc 

    # Create a new Spark session from the current Spark session (with isolated SQL configurations).
    # The new Spark session is sharing the underlying SparkContext and cached data,
    # but the temporary tables and registered functions are isolated.
    def createNewSparkSession(self, currentSparkSession):
        self.spark = currentSparkSession.newSession()
        self._init()
        return self.spark

    def getSparkSession(self):
        return self.spark

    def getSparkSessionConf(self):
        return self.spark.conf

    def getSparkContext(self):
        return self.sc

    def getSparkContextConf(self):
        return self.scConf

    def getSparkContextConfAll(self):
        return self.scConf.getAll()

    def setSparkContextConfAll(self, properties):
        # Properties example: { 'spark.executor.memory' : '4g', 'spark.app.name' : 'Spark Updated Conf', 'spark.executor.cores': '4',  'spark.cores.max': '4'}
        self.scConf = self.scConf.setAll(properties) # or self.scConf = self.spark.sparkContext._conf.setAll()

    # Stop (clears) the active SparkSession for current thread.
    #def stopSparkSession(self):
    #    return self.spark.clearActiveSession()

    # Stop the underlying SparkContext.
    def stopSparkContext(self):
        self.spark.stop() # Or self.sc.stop()

    # Returns the active SparkSession for the current thread, returned by the builder.
    #def getActiveSparkSession(self):
    #    return self.spark.getActiveSession()

    # Returns the default SparkSession that is returned by the builder.
    #def getDefaultSession(self):
    #    return self.spark.getDefaultSession()